CN104937833A - Controlling electrical converter - Google Patents
Controlling electrical converter Download PDFInfo
- Publication number
- CN104937833A CN104937833A CN201380059764.1A CN201380059764A CN104937833A CN 104937833 A CN104937833 A CN 104937833A CN 201380059764 A CN201380059764 A CN 201380059764A CN 104937833 A CN104937833 A CN 104937833A
- Authority
- CN
- China
- Prior art keywords
- gradient
- variable
- calculating
- optimized
- matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02M—APPARATUS FOR CONVERSION BETWEEN AC AND AC, BETWEEN AC AND DC, OR BETWEEN DC AND DC, AND FOR USE WITH MAINS OR SIMILAR POWER SUPPLY SYSTEMS; CONVERSION OF DC OR AC INPUT POWER INTO SURGE OUTPUT POWER; CONTROL OR REGULATION THEREOF
- H02M1/00—Details of apparatus for conversion
- H02M1/08—Circuits specially adapted for the generation of control voltages for semiconductor devices incorporated in static converters
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02P—CONTROL OR REGULATION OF ELECTRIC MOTORS, ELECTRIC GENERATORS OR DYNAMO-ELECTRIC CONVERTERS; CONTROLLING TRANSFORMERS, REACTORS OR CHOKE COILS
- H02P21/00—Arrangements or methods for the control of electric machines by vector control, e.g. by control of field orientation
Landscapes
- Engineering & Computer Science (AREA)
- Power Engineering (AREA)
- Control Of Ac Motors In General (AREA)
- Supply And Distribution Of Alternating Current (AREA)
- Feedback Control In General (AREA)
- Dc-Dc Converters (AREA)
Abstract
A method for controlling an electrical converter (12) comprises the steps of: receiving an actual electrical quantity (20) relating to the electrical converter (12) and a reference quantity (22); determining a future state of the electrical converter (12) by minimizing an objective function based on the actual electrical quantity and the reference quantity as initial optimization variables; and determining the next switching state for the electrical converter (12) from the future state of the electrical converter (12). The objective function is iteratively optimized by: calculating optimized unconstrained optimization variables based on computing a gradient of the objective function with respect to optimization variables; and calculating optimization variables for a next iteration step by projecting the unconstrained optimization variables on constraints. The computation of the gradient and/or the projection is performed in parallel in more than one computing unit (44, 50, 70).
Description
Technical field
The present invention relates to method for controlling electric transducer and controller and electric transducer.
Background technology
Electric transducer is used for converting first electric current (such as, DC electric current or AC electric current) with first frequency second electric current with other frequency to, such as other DC or AC electric current.
Especially, Power electronic converter (it can be connected to motor maybe can be interconnected to electrical network) generally includes a large amount of power semiconductor, must switch and control these power semiconductors and generate the output current of expectation.Except the minimizing of the difference between reference quantity and actual determined amounts (such as flux (flux) and torque), the important goal of control can be the short response time of controller, the low harmonics distortion of the electric current of generation and low switching losses.
Some or all be formulated as target functions in these targets, the possible to-be of its reception electric transducer is as input variable and output must be minimized the value at cost reaching target mentioned above.
To-be (such as, it comprises following on off sequence, non-incoming current, following voltage, etc.) then can be used for the next on off state determining to be applied to transducer.
Exemplarily, consider that model prediction pulse mode controls (MP3C), it is described in more detail in EP 2 469 692 A1.MP3C is the method based on Model Predictive Control (MPC), and it is in conjunction with direct torque control (DTC) and optimize pulse mode (OPP) these two classics and the good advantage for controlling the torque of motor and the method for flux in middle pressure driving application of establishing.Result can be control and modulation strategy, and it produces the nearly optimal ratio of response time very short between transient period, the quick harmonic current distortions at the every switching frequency of steady state operation suppressed and cause owing to using OPP of interference.Method may be applicable to the Power electronic converter being connected to motor and electrical network.
MP3C is from the low harmonic current distortion using precalculated optimization pulse mode (it can be adaptive to realize optimum performance between transient period online) to inherit it.Adopt its primitive form, method needs to determine must correcting of the switching time of precalculated pulse mode in the solution of each sampling instant quadratic programming (QP).But standard QP solver has high calculation requirement usually.
MP3C just needs an example of the accurate transient solution of optimization problem.Another application example can be control Power electronic converter, and its middle controller calculates modulation index.This available controller cascade with external voltage control ring and interior current regulator (both based on MPC) realizes.Current controller can need the solution of QP and voltage controller can to multinomial optimization problem.
Combine Decision and Control (CDC/CCC 2009) in 2009 that the 28th Chinese Control Conference in 2009 hold, the 48th IEEE conference agenda 7387-7393 page in 2010, S.Richter, C. N. Jones and M. Morari article " Real-time input-constrained MPC using fast gradient methods(uses the real-time input constraint MPC of Fast Field method) " Fast Field method for calculating optimum cost value to a class target function is shown.
Summary of the invention
Target of the present invention is to provide the controller and control method with short response time, and it produces low harmonics distortion and generates low switching losses.
This target is realized by the purport of independent claims.Other one exemplary embodiment from dependent claims and following description apparent.
Aspect of the present invention relates to the method for controlling electric transducer (such as middle pressure converter, it can be DC to AC, AC to DC or AC to AC transducer).The method can realize in the controller of electric transducer, such as FPGA or DSP.Electric transducer can be the part of the system comprising electric transducer and motor (such as electro-motor or generator).But electric transducer also can make the first electrical network be connected with the second electrical network.
According to embodiments of the invention, method comprises the following steps: the actual electricity and the reference quantity that receive electric transducer; At least one possibility to-be (or multiple possibility to-be) by making the minimization of object function determine electric transducer based on actual electricity and reference quantity; And determine the next on off state for electric transducer from the possible to-be of electric transducer.
Control method can be ring control methods, it can make one or more actual electricity (such as, actual output current, actual output voltage, real fluxes and/or actual torque) and reference quantity is (such as, with reference to output current, reference output voltage, reference flux, with reference to actual torque) between minimize variability, this one or more actual electricity can be measured or indirectly determine from the amount measured.
In addition, control method can make other control objectives minimize as described above.All control objectives can be encoded in target function, and the amount as predicted state and input is mapped to value at cost as optimized variable by this target function.Electricity can be actual and the other amount of the to-be of reference quantity and transducer and/or system, and it such as can comprise the following voltage of semiconductor switch, electric current, flux, torque, on off state, etc.
Target function and/or to-be can based on the physical models that can obtain from the setting of electric transducer and/or whole electric system.
In each control cycle, control method determines the next on off state of the semiconductor switch for electric transducer by selecting obedience to have the to-be of the target function of the best (such as, minimum) value at cost.
Target function is by following and iteration optimization: the unconstrained optimization variable carrying out calculation optimization based on the gradient about optimized variable calculating target function; And by unconstrained optimization variable drop being calculated in constraint the optimized variable for next iterative step.Such as, target function can calculate by the Fast Field method such as described in the article of the people such as Richter.
The calculating of gradient and/or projection executed in parallel in more than a computing unit.Computing unit can be the unit of FPGA or the core of part or polycaryon processor.Polycaryon processor can comprise more than a processor on a single die or can comprise some processors on different chips.
Such as, gradient calculating can in more than a computing unit executed in parallel.Alternatively or in addition, gradient calculating can with projection executed in parallel.The calculating of gradient and/or projection can perform and/or can pipelining in parallel thread.
According to embodiments of the invention, iteration comprises further: divided by optimized variable in groups, makes often to organize optimized variable and can organize to separate with other and project.Electricity about the electric phase place grouping of electric transducer and/or can be divided into physical similarity or equal quantities, such as, be divided into electric current, voltage, on off state, flux etc.In addition, grouping can be carried out about projection, make each group can independent of each other or the projection that is separated from each other.Grouping also can be carried out about the submatrix of the matrix for compute gradient, and wherein these submatrixs are haply independent of (that is, can only there is little matrix entries in submatrix outside) each other.
The gradient of some computing units of controller for each group of executed in parallel unconstrained optimization variable and/or the calculating of projection can be used.Adopt in such a way, the intercommunication between computing unit can be almost optional.
According to embodiments of the invention, optimize unconstrained optimization variable by the gradient that is negative and/or suitably calibration of target function is added to optimized variable to calculate.That is, optimized variable is by optimizing towards best stepping in antigradient direction.
According to embodiments of the invention, iteration comprises to calibrate constrained optimization variable by scaling factor.Such as, the difference between the optimized variable of the optimized variable that projection (that is, retraining) is optimized and iterative step is before calibrated by scaling factor, with the calculating of the optimized variable of adjusting and optimizing, as described below.Calibration can controller more than a computing unit in perform.Calibration can with gradient calculation and at least one executed in parallel in projecting.
According to embodiments of the invention, the gradient of target function comprises Matrix Multiplication with the vector of optimized variable.When target function forms quadratic programming problem, situation can be like this.
According to embodiments of the invention, the executed in parallel in more than a computing unit that is multiplied of the entry of matrix column and optimized variable.This can be suitable for performing iteration in FPGA, because calculate by calculating and streamlined each row executed in parallel of matrix.
According to embodiments of the invention, the executed in parallel in more than a computing unit that is multiplied of the entry of the row of matrix and optimized variable.This can be suitable for executed in parallel iteration in different threads, because can only there is little inter-thread communication.
According to embodiments of the invention, method can comprise the following steps: by the sequence making the minimization of object function determine following on off state; And use from the first to-be of the sequence of this following on off state as the next on off state that will be applied to electric transducer.Control method can based on mobile time domain (moving horizon), namely not only determines the next one in each cycle but also determines the to-be of many (time domain) following control cycle.
Other aspect of the present invention relates to the controller for electric transducer, wherein this controller be suitable for perform as above with method described below.It must be understood that, as above with the feature of method described below can be as above with the feature of the controller of following description, and vice versa.
Controller can comprise computing unit, and it is for executed in parallel method step mentioned above.Controller can comprise FPGA, CPU and/or GPU, for providing computing unit.
Can realize at FPGA or based in the control system of multinuclear the part of the controller of Optimization Solution, to its programming with admit actual amount (such as, systematic survey) and (such as to-be) other control problem parameters as input, perform required calculating and return description control inputs by the variable handled.
According to embodiments of the invention, controller comprises field programmable gate array (FPGA), it comprise following at least one: at least one matrix multiplication unit, for making optimized variable and matrix multiple; At least one projecting cell, for the unconstrained optimization variable that projects; At least one scaling unit, is calibrated constrained optimization variable for being come by scaling factor.
Calculating can at least two matrix multiplication unit, at least two projecting cells and/or at least two scaling unit executed in parallel.Calculating can be made their executed in parallel in matrix multiplication unit, projecting cell and/or scaling unit by streamlined.
Optimization method based on gradient can realize in FPGA, at microsecond range to optimization problem.The calculating strength realized in conjunction with algorithm advantage and the FPGA of the algorithm based on gradient in MPC application can allow to solve the quadratic programming problem of the MP3C with large prediction time domain best, this can cause the control performance improved, such as torque and stator flux (stator flux) can keep close to they references between transient period, and harmonic current distortions is in its physics minimum value simultaneously.
Controller can comprise mixed architecture, the combination (such as on identical chips) of such as CPU and FPGA.Such as, in the enterprising row matrix-vector multiplication of FPGA, and then perform on CPU (may be complicated) projection.
If control system is combined with FPGA, CPU or DSP of controller is by being set up the required data of write to arrange control problem at predefine memory bit.FPGA can directly measure/estimate actual amount and to the MPC problem solving of gained, or can receive these values from DSP/CPU.Similarly, optimal control actions directly can be sent to modulator by FPGA, or result of calculation is returned DSP/CPU and be used for further process.
According to embodiments of the invention, the gradient of target function comprises vector portion, and it is based on the matrix equation of actual amount and/or reference quantity.Compute vector part before matrix multiplication unit is used in and calculates unconstrained optimization variable.Vector portion can calculate once from actual and/or reference quantity in each cycle.In order to save independent computing unit, the matrix multiplication unit for the matrix-vector multiplication of calculating target function also can the matrix-vector multiplication of compute vector part.
According to embodiments of the invention, controller comprises polycaryon processor, its middle controller be suitable for polycaryon processor more than a core in optimized variable group executed in parallel gradient calculation and/or projection.Can perform in a thread (performing in its in core) the gradient calculation of one group of optimized variable, projection and/or calibration.
Other aspect of the present invention relates to electric transducer, it comprise as above with controller described below.
These and other aspects of the present invention will be obvious from embodiment described below and illustrate with reference to them.
Accompanying drawing explanation
Purport of the present invention will be explained in further detail with reference to illustrated one exemplary embodiment in the accompanying drawings in following text, wherein:
Fig. 1 schematically illustrates electric transducer according to an embodiment of the invention.
Fig. 2 illustrates according to embodiments of the invention for controlling the flow chart of the method for transducer.
Fig. 3 schematically illustrates the controller according to embodiments of the invention with FPGA.
Fig. 4 illustrates the figure of the operation of the FPGA of key-drawing 3.
Fig. 5 schematically illustrates the controller according to embodiments of the invention with polycaryon processor.
Fig. 6 illustrates the figure of the operation of the controller of key-drawing 5.
In principle, identical parts provide identical label in the drawings.
Embodiment
Fig. 1 illustrates system 10, and it comprises electric transducer 12, and this electric transducer 12 can be the transducer of Power electronic converter (as modular multistage transducer) or any other type.In example shown in Figure 1, transducer 12 converts three-phase input current to three-phase output current.Such as, but the configuration of any other type is possible, DC to AC transducer or AC to DC transducer.
System 10 comprises load 14, and it is by electric transducer 12(such as motor 14 or other electrical network 14) power supply.
In order to switching current, transducer 12 comprises a large amount of semiconductor switch, and it is controlled by controller 16.
Generally, controller 16 calculates the next on off state 18 that must be applied to transducer 12.Next on off state 18 is determined from the actual electricity 20 of system 10 and one or more reference quantity 22 based on minimizing of the target function of encoding to the control objectives of electric system 10.
Actual electricity 20 can comprise to be measured and/or estimates and can come from electric transducer 12, load (machine or electrical network) 14 and/or come from the electric network of powering to electric transducer 12.
Fig. 2 illustrates the method for controlling transducer 12.
In step 30, controller 16 receives one or more actual electricity (such as from the measurement mechanism being attached to controller 16 or load 14) and one or more reference quantity 22.
Actual electricity 20 can comprise the electric current (such as in the output of transducer 12 or transducer 12 inside) of measurement, the flux in the measuring voltage, motor of the output of transducer 12 or converter inside measurement and torque.
Reference quantity 22 can comprise reference current, reference voltage, reference flux etc.
In the step 32, controller 16 determines one or more to-bes of transducer 12 from actual amount 20 and reference quantity 22, and it can comprise following on off state and/or following electricity.
Such as, controller 12 is by generating the following on off state sequence of possibility and obtaining following electricity thus to determine following electricity sequence.In this step, the model of transducer 12 and/or load 14 can be used for calculating following electricity.
As another example, controller 16 can determine the precalculated pulse mode as following on off state, as described in EP 2 469 692 A1.
In step 34, controller 16 makes objective function optimization (such as, minimizing it) for determining to have the to-be of optimum cost value.Generally, target function is function to-be being mapped to value at cost.
Such as, target function is by also making the harmonic distortion in input and/or output current minimize and/or making the minimize variability between following electricity and reference electric quantity by making the switching loss of transducer 12 minimize.
Target function can obtain from Operation system setting and physical law.
In step 36, next on off state obtains from the to-be with optimization object function.
Such as, controller 16 can comprise modulator, and it receives reference voltage based on the to-be optimized, and other controller determines next on off state from this reference voltage.
As other example, when to-be comprises to-be sequence or series, mobile time domain policy can be used and only can select the on off state of next to-be.
Below, the optimization of target function f (z) will be explained in more detail in step 34.In many cases, target function f (z) is the polynomial function of its input variable z, namely can be the virtual condition of system 10 and the value of to-be.
Optimization problem mentioned above can be formulated as the continuous optimization problems with following form:
Wherein optimized variable
, target function f (z) and constraint set Z=Z
1× Z
2×× Z
n, the cartesian product of itself and simple set is similar.If the projection (vide infra) of the point on collection be computationally lightweight compared with original optimization problem, this collection can be described as simple.Simple set can be considered one group of constraint interlaced with each other.
Such as, due to the form of physical law (such as, Kirchhoff's law), target function normally secondary, it obeys the double optimization problem with following form:
Wherein
symmetric positive semidefinite matrix,
it is vector.
As will be become important below, the gradient of this secondary target function
provided by 2Hz+g.
In the article (mentioning at the beginning) of the people such as Richter, propose to be used for two possibility methods to the problem solving for MPC application above: gradient and Fast Field method.
Fast Field method is here reaffirmed as algorithm.This algorithm needs initial point
,
as input variable, iteration number i
max, Lipschitz constant L and scale factor β
0..., β
imax-1.
For i=0,
available
Terminate
Algorithm makes variable z optimize about target function f (z).
In the first row in iteration loop, algorithm is from the optimized variable y calculated before during iterative step before
icalculate unconstrained optimization variable
.This passes through in gradient
opposite direction on stepping and carrying out.
In the second row in iteration loop, algorithm is by unconstrained optimization variable
project to constrained optimization variable z
i+1on.
In the third line in iteration loop, algorithm is by scaling factor β
0..., β
imax-1come constrained optimization variable z
i+1calibrate with the optimized variable y calculated for next iterative step
i+1.By calibrating optimized variable, recoverable gradient
opposite direction on step.
Symbol
projection on instruction constraint set Z.Such as, some in to-be are confined to some boundary or only can comprise centrifugal pump.In these cases, project
by making without binding occurrence
be confined to boundary and/or make them be confined to binding occurrence z by they being arranged to next permission centrifugal pump.
In MPC problem, optimized variable z represents the control inputs after using state renewal equation elimination status predication usually.Collection Z
ithen the constraint to control inputs in prediction time domain can be represented.When existence retrains, the elimination of status predication can produce constraint set Z
i, it is no longer simple.But the nested solution via the optimization problem only with simple set utilizes gradient method to MPC problem solving, this can remain possible.Therefore, the method provided also can be used for the solution with state constraint or the general MPC problem to the soft-constraint of state/input.
In addition, note in MPC is arranged, the function of vectorial g current measurement typically (or estimation) actual amount and reference quantity, and therefore need to upgrade in each cycle of control ring.
The control method that above-described use optimized algorithm carrys out optimization object function can adopt following manner at FPGA(see Fig. 3) or the middle realization of multinuclear controller (see Fig. 5).
Optimized variable z is divided into N number of group,
, wherein
,
.The size n of group
ican be equal.But the large I of all groups is unequal, this is possible.
Preferably, all elements of a group is under the jurisdiction of a simple set
and other groupings whole collection of optimized variable (in the such as single element or a group) are possible.
Following discussion will focus on has secondary target function
quadratic programming problem on.
Use piecemeal, the Matrix-Vector multiplication calculated for gradient step may be prescribed as
Wherein each c
ithe n of oriental matrix M=I-2/LH
ithe collection of individual column vector.
This piecemeal is useful for the realization in FPGA.
Fig. 3 illustrates the controller 16 with FPGA 38, this FPGA 38 has many computing units of the algorithm that realization illustrates above, i.e. data input cell 40, matrix unit 42, one or more matrix multiplication unit 44, add tree unit 46, buffer cell 48, many projecting cells 50 and many beta unit 52 and data outputting unit 54.
The operation of FPGA 40 will be explained about Fig. 4, and this Fig. 4 illustrates the Gantt chart in the cards (Gantt chart) of the Fast Field method (algorithm) on FPGA.
Horizontal line represents computing unit: data input cell 40, matrix unit 42, Matrix-Vector multiplication unit 44 and add tree unit 46, projecting cell 50, beta unit 52 and data outputting unit 54.Horizontal axis repre-sents time, each indentation indicates a clock cycle.Box indicates the use of corresponding computing unit, and arrow manifests the transfer of data from a computing unit to another.
In section 60, the vector portion of the gradient of calculating target function f (z)
.The component of gradient
depend on the reality of transducer 12 and/or load 14 and reference quantity and each cycle of controller 16 must the component of compute gradient
(wherein performing all iteration being used for optimization object function).
Usually,
based on matrix equation and matrix multiplication unit 44 can be used for calculate
.In section 60, for
corresponding matrix be loaded in matrix multiplication unit 44 by matrix unit 42.In addition, input data cell load 40 reality and reference quantity are loaded in matrix multiplication unit 44.
Same in section 60, the initial value of optimized variable z can be loaded in matrix multiplication unit 44.
In section 62, the first iteration is performed by matrix multiplication unit 44 and calculates c
jz
jand perform result to be added to register or buffer cell 48(by add tree unit 46
stored therein, alternatively,
can be stored in matrix multiplication unit 44).
If only use a matrix multiplication unit 44, in each clock cycle, by one group of z
ibe delivered to this unit, make after N number of clock cycle, obtain
.Be used in each clock cycle more than a matrix multiplication unit 44 and calculate more than a group.
At the end of section 62, first group of optimized variable z
iproject to the upper and result of corresponding constraint by constraint projecting cell 50 to be calibrated by beta unit 52.
Calibrating first group of z
itime, matrix multiplication unit 44 can start the iteration performed then again.In these computing intervals, optimized variable group is then parallel to be projected and to calibrate and is then delivered to matrix multiplication unit 44.
If the element of superior vector z is grouped make each group of z
iindependently can be projected, make group z
istreamlined improved by computing unit the computing unit of FPGA 38 utilization and thus on FPGA, save space and on the processing time, there is no large delay, this is possible.
In section 64a, 64b etc., perform other iteration.Reaching i
maxduring the limit of iteration, in section 66, data outputting unit 54 reads result.Note when mobile time domain, data outputting unit only can need the first element reading optimized variable z, and it comprises the next to-be of transducer 12.
As alternative, Matrix-Vector multiplication can not be implemented as row-vector multiplication as row-scalar multilication:
Wherein r
ithe n of M
ithe collection of individual row.Because lower than in row scale method of intercore communication amount, OK-vector formization can ask multi-core platform especially.
Fig. 5 illustrates controller 16, and it comprises multiple single core processor 70, and they can be considered that the computing unit 40 to 54 of computing unit 70(and the FPGA 38 of the other types of controller 16 is compared).
Fig. 6 illustrates the figure of functional module and the buffer performed on the controller of Fig. 5.Data flow between arrow indicator collet.
In a first step, the initial value for optimized variable is stored in input buffer 72.As indicated in figure 6, optimized variable y
ibe divided into three groups.But the group of any other quantity is also possible.
In addition, the entry of matrix M and vector
be stored in matrix buffer 74.
After that, executed in parallel gradient and projection module 76 in core 70.With optimized variable z
ithe row of matrix buffer 74 of respective sets association load from matrix buffer 74 and be multiplied with all variablees from buffer 72 and be projected the optimized variable z of calculation optimization
i, it is stored in buffer 78.
Calculating relevant variable z
iafter, in each core 70, perform scaling module 80, it calculates the optimized variable y for the calibration of next iterative step
i.In this step, from the optimized variable z of last step
iload from buffer 82.
At the end of iteration, all core must wait for that thread stops that 86 make the execution of next iteration synchronous.
Buffer 72,84 can be embodied as read buffer 72 and write buffer 84, and it has two pointers, and it exchanges in thread stop place and avoids unnecessary copying.
Another embodiment is the realization for the more method based on gradient of the solution of general optimization problem.F (z) has gradient
general differentiable target function, instead of secondary target function f (z).(fast) gradient method that non-convex differentiable target function adopts does not ensure converge to global minimum and ensure on the contrary to converge to local minimum.Although the streamlined of above-described computing unit 50 to 54 should be suitable for the structure of problem at the moment, the realization on FPGA is still possible.For simplicity, we suppose z
ibe scalar, notice that it is simple for making the discussion presented expand to non-scalar piecemeal.
If f (z) is d multinomial, gradient
by the n with maximum d-1 time
zindividual multinomial composition.Illustratively property example, considers the multinomial target function had as Gradient
:
Such as, Parallel Implementation is undertaken by compute gradient valuation line by line.
On FPGA or polycaryon processor, save space in order to speed-up computation, evaluation block can reuse multiplication result, such as, by the polynary Huo Na scheme of application.
In order to illustrate, consider the multinomial of the example previously presented
.The simple realization of multinomial valuation is by needs 6 multipliers.On the contrary, by using simple Huo Na scheme, it can be rewritten as by we:
, therefore cause 4 multiplication.In addition, item 2z
1only can calculate once, thus be reduced to sum 3 multiplication.Except the lower quantity operation needed, Huo Na scheme also allows to reuse identical multiplier and adder unit.
Consider the specific specificity of problem, the method can adjust further and make the intermediate object program of Huo Na scheme can be reused for multinomial valuation then.
Again consider example: the intermediate object program for the valuation of the first polynomial Huo Na scheme is
.Be not also develop Huo Na scheme completely to the second multinomial, solution can by reusing
, add
and calculate simply
and adopt single multiplier.
Such method can be adopted for general differentiable target function.
Although the present invention has illustrated in detail and described in accompanying drawing and aforementioned description, such diagram and describing will have been thought illustrative or exemplary instead of restrictive; The invention is not restricted to disclosed embodiment.To other versions of disclosed embodiment can by those skilled in that art from accompanying drawing, the study of claim that is open and that enclose to be come in the present invention of practice calls protection understand and implement.In the claims, word " comprises " does not get rid of other elements or step, and indefinite article " " does not get rid of majority.Single processor or controller or other unit can meet the function of the some projects enumerated in the claims.The only fact that some measure is recorded in mutually different dependent claims does not indicate the combination of these measures cannot be used to realize advantage.Any label in the claims should not be interpreted as limited field.
Claims (15)
1., for controlling a method for electric transducer (12), described method comprises step:
Reception relates to actual electricity (20) and the reference quantity (22) of described electric transducer (12);
The possible to-be by making the minimization of object function determine described electric transducer (12) based on described actual electricity and described reference quantity; And
The next on off state for described electric transducer (12) is determined from the possible to-be of described electric transducer (12);
Wherein said target function is by following and iteration optimization:
The unconstrained optimization variable of calculation optimization is carried out based on the gradient calculating described target function about optimized variable; And
By described unconstrained optimization variable drop being calculated in constraint the optimized variable for next iterative step;
The calculating of wherein said gradient and/or described in be projected in more than executed in parallel in a computing unit (44,50,70).
2. the method for claim 1,
The calculating of wherein said gradient is more than executed in parallel in a computing unit (44,70).
3. method as claimed in claim 1 or 2,
The calculating of wherein said gradient and described projection executed in parallel.
4. the method as described in any one in claim 1-3,
Wherein said iteration comprises:
Described optimized variable is divided in groups, makes often to organize optimized variable and can organize to separate with other and project;
The calculating of wherein said gradient and/or the projection of described unconstrained optimization variable use some computing units (44,50, the 70) executed in parallel of described controller (16).
5. the method as described in any one in claim 1-4,
Wherein said iteration comprises to be come described constrained optimization variable calibration by scaling factor;
Wherein said at least one executed in parallel be targeted at more than performing in a computing unit (52,70) and/or in wherein said calibration and described gradient calculation and described projection.
6. the method as described in any one in claim 1-5,
The gradient of wherein said target function comprises Matrix Multiplication with the vector of optimized variable.
7. method as claimed in claim 6,
The entry of wherein said matrix column is multiplied with optimized variable executed in parallel in more than a computing unit.
8. method as claimed in claim 6,
Wherein the entry of the row of matrix is multiplied with optimized variable executed in parallel in more than a computing unit.
9. the method as described in any one in claim 1-8,
The unconstrained optimization variable wherein optimized calculates by adding the negative gradient of described target function to described optimized variable.
10. the method as described in any one in claim 1-9, it is further comprising the steps:
By the sequence making described the minimization of object function determine following on off state; And
Use from the first to-be of the sequence of described following on off state as the next on off state that will be applied to described electric transducer.
11. 1 kinds of controllers for electric transducer (12) (16), wherein said controller (16) is suitable for performing the method as described in any one in claim 1 to 10.
12. controllers (16) as claimed in claim 11, it has FPGA(38), described FPGA comprise following at least one:
At least one matrix multiplication unit (44), for making described optimized variable and matrix multiple;
At least one projecting cell (50), for the described unconstrained optimization variable that projects;
At least one scaling unit (52), for coming by scaling factor described constrained optimization variable calibration.
13. controllers (16) as claimed in claim 12,
The gradient of wherein said target function comprises vector portion, and it is based on the matrix equation of described actual amount and/or described reference quantity;
Wherein said matrix multiplication unit (44) for calculating described vector portion before the described unconstrained optimization variable of calculating.
14. controllers (16) as claimed in claim 11, it comprises:
Polycaryon processor,
Wherein said controller be suitable for described polycaryon processor more than a core (70) in concurrently described gradient calculation and/or described projection are performed to optimized variable group.
15. 1 kinds of electric transducers (12), it comprises the controller (16) as described in any one in claim 11 to 14.
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| EP12192811.3A EP2733842B1 (en) | 2012-11-15 | 2012-11-15 | Controlling an electrical converter |
| EP12192811.3 | 2012-11-15 | ||
| PCT/EP2013/073798 WO2014076167A2 (en) | 2012-11-15 | 2013-11-14 | Controlling an electrical converter |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104937833A true CN104937833A (en) | 2015-09-23 |
| CN104937833B CN104937833B (en) | 2018-04-27 |
Family
ID=47500888
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201380059764.1A Active CN104937833B (en) | 2012-11-15 | 2013-11-14 | Control electric transducer |
Country Status (4)
| Country | Link |
|---|---|
| US (1) | US20150249381A1 (en) |
| EP (1) | EP2733842B1 (en) |
| CN (1) | CN104937833B (en) |
| WO (1) | WO2014076167A2 (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109995257A (en) * | 2017-12-07 | 2019-07-09 | Abb瑞士股份有限公司 | The method and converter of control and modulating converter |
| CN112868174A (en) * | 2018-10-17 | 2021-05-28 | 保时捷股份公司 | Control of a modular multipoint serial-to-parallel converter (MMSPC) by means of a switching table and its continuous background optimization |
| CN113437922A (en) * | 2021-07-27 | 2021-09-24 | 上海莘汭驱动技术有限公司 | Driving control method and system for limited-angle torque motor |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2016177535A1 (en) | 2015-05-05 | 2016-11-10 | Abb Schweiz Ag | Hybrid control method for an electrical converter |
| CN117933327A (en) * | 2017-04-21 | 2024-04-26 | 上海寒武纪信息科技有限公司 | Processing device, processing method, chip and electronic device |
| CN113093542B (en) * | 2021-03-31 | 2022-08-12 | 吉林大学 | Motor torque optimization finite set prediction control parallel computing method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6373219B1 (en) * | 1999-06-22 | 2002-04-16 | Hitachi, Ltd. | Motor control system and motor control method |
| EP1670135A1 (en) * | 2004-12-10 | 2006-06-14 | Abb Research Ltd. | Method of operating a rotary electrical machine |
| CN101388641A (en) * | 2007-09-10 | 2009-03-18 | Abb研究有限公司 | Method for operating an electric rotary machine |
| US20120068641A1 (en) * | 2010-09-21 | 2012-03-22 | Denso Corporation | Control device for electric rotating machine |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6590370B1 (en) * | 2002-10-01 | 2003-07-08 | Mti Microfuel Cells Inc. | Switching DC-DC power converter and battery charger for use with direct oxidation fuel cell power source |
| WO2004042889A1 (en) * | 2002-11-04 | 2004-05-21 | Jovan Bebic | Hybrid power flow controller and method |
| CA2576778C (en) * | 2006-02-07 | 2014-09-02 | Xinping Huang | Self-calibrating multi-port circuit and method |
| KR100886194B1 (en) * | 2007-06-08 | 2009-02-27 | 한국전기연구원 | Controller of double-fed induction generator |
| WO2011098100A1 (en) * | 2010-02-11 | 2011-08-18 | Siemens Aktiengesellschaft | Control of a modular converter having distributed energy stores by means of an observer for the currents and by means of an estimating unit for the intermediate circuit energy |
| EP2469692B1 (en) | 2010-12-24 | 2019-06-12 | ABB Research Ltd. | Method for controlling a converter |
| US9390370B2 (en) * | 2012-08-28 | 2016-07-12 | International Business Machines Corporation | Training deep neural network acoustic models using distributed hessian-free optimization |
-
2012
- 2012-11-15 EP EP12192811.3A patent/EP2733842B1/en active Active
-
2013
- 2013-11-14 WO PCT/EP2013/073798 patent/WO2014076167A2/en active Application Filing
- 2013-11-14 CN CN201380059764.1A patent/CN104937833B/en active Active
-
2015
- 2015-05-14 US US14/712,444 patent/US20150249381A1/en not_active Abandoned
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6373219B1 (en) * | 1999-06-22 | 2002-04-16 | Hitachi, Ltd. | Motor control system and motor control method |
| EP1670135A1 (en) * | 2004-12-10 | 2006-06-14 | Abb Research Ltd. | Method of operating a rotary electrical machine |
| CN101388641A (en) * | 2007-09-10 | 2009-03-18 | Abb研究有限公司 | Method for operating an electric rotary machine |
| US20120068641A1 (en) * | 2010-09-21 | 2012-03-22 | Denso Corporation | Control device for electric rotating machine |
Non-Patent Citations (1)
| Title |
|---|
| STEFAN RICHTER ET AL.: "High-speed Online MPC Based on a Fast Gradient Method Applied to Power Converter Control", 《2010 AMERICAN CONTROL CONFERENCE MARRIOTT WATERFRONT, BALTIMORE,MD,USA》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109995257A (en) * | 2017-12-07 | 2019-07-09 | Abb瑞士股份有限公司 | The method and converter of control and modulating converter |
| CN109995257B (en) * | 2017-12-07 | 2022-07-26 | Abb瑞士股份有限公司 | Method for controlling and modulating a converter and converter |
| CN112868174A (en) * | 2018-10-17 | 2021-05-28 | 保时捷股份公司 | Control of a modular multipoint serial-to-parallel converter (MMSPC) by means of a switching table and its continuous background optimization |
| CN113437922A (en) * | 2021-07-27 | 2021-09-24 | 上海莘汭驱动技术有限公司 | Driving control method and system for limited-angle torque motor |
Also Published As
| Publication number | Publication date |
|---|---|
| US20150249381A1 (en) | 2015-09-03 |
| CN104937833B (en) | 2018-04-27 |
| EP2733842B1 (en) | 2018-07-04 |
| WO2014076167A2 (en) | 2014-05-22 |
| EP2733842A1 (en) | 2014-05-21 |
| WO2014076167A3 (en) | 2014-12-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US9964980B2 (en) | Method and apparatus for optimal power flow with voltage stability for large-scale electric power systems | |
| CN104937833A (en) | Controlling electrical converter | |
| Bai et al. | Semi-definite programming-based method for security-constrained unit commitment with operational and optimal power flow constraints | |
| US9099866B2 (en) | Apparatus, methods and systems for parallel power flow calculation and power system simulation | |
| Oliveira et al. | Short term hydroelectric scheduling combining network flow and interior point approaches | |
| Jubril et al. | Solving multi-objective economic dispatch problem via semidefinite programming | |
| Hesamzadeh et al. | Computation of extremal-Nash equilibria in a wholesale power market using a single-stage MILP | |
| Kolen et al. | Enabling the analysis of emergent behavior in future electrical distribution systems using agent‐based modeling and simulation | |
| Chalangar et al. | Methods for the accurate real-time simulation of high-frequency power converters | |
| Geth et al. | Convex power flow models for scalable electricity market modelling | |
| CA2677384A1 (en) | Apparatus, methods and systems for parallel power flow calculation and power system simulation | |
| Boéchat et al. | An architecture for solving quadratic programs with the fast gradient method on a Field Programmable Gate Array | |
| Ma et al. | Real-time simulation of power system electromagnetic transients on FPGA using adaptive mixed-precision calculations | |
| CN109615151A (en) | A kind of prediction technique, device and the medium of the double optimizations of load energy storage | |
| Rico-Hernández et al. | Analysis of electrical networks using fine-grained techniques of parallel processing based on OpenMP | |
| Louie et al. | Hierarchical multiobjective optimization for independent system operators (ISOs) in electricity markets | |
| Sturtz et al. | Accelerating the neural network controller embedded implementation on FPGA with novel dropout techniques for a solar inverter | |
| Khaitan et al. | TDPSS: a scalable time domain power system simulator for dynamic security assessment | |
| Kulisz et al. | A hardware implementation of the PID algorithm using floating-point arithmetic | |
| Novak et al. | Implementation of mixed-integer programming on embedded system | |
| Vyncke et al. | Simulation-based weight factor selection and FPGA prediction core implementation for finite-set model based predictive control of power electronics | |
| CN111105100B (en) | Neural network-based optimization method and system for multi-microgrid scheduling mechanism | |
| Marufuzzaman et al. | A high speed current dq PI controller for PMSM drive | |
| Zimmer et al. | Implementation of a modelica library for energy management based on economic models | |
| Zenor et al. | Efficient real-time simulation of linear differential equations arising from simulation of electronic power systems |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20180509 Address after: Baden, Switzerland Patentee after: ABB TECHNOLOGY LTD. Address before: Zurich Patentee before: ABB T & D Technology Ltd. |
|
| TR01 | Transfer of patent right |